Differential output fitting has one failure mode, and it is structural. You implement the general algorithm, run the same molecules through both libraries, cluster the mismatches, infer the missing rule, patch, then grow the corpus and repeat [7]. Each growth round tends to surface a new mismatch family, each family earns a new branch, and the corpus quietly becomes the specification the implementation is written against [8]. That is why a pass rate stops being informative near the end. As the post argues, 99.9% does not tell you whether the residual is ten stragglers or a semantic family nobody has modelled [9], and no corpus size retires the risk that production exercises a combination the corpus never held [10].
The alternative COSMolKit describes is to derive compatibility-critical chemistry from the pinned upstream source, then use the differential corpora to audit that port rather than to teach it [6]. Its worked example is two implementations that both report 1,000,000 of 1,000,000: one repaired until the outputs agreed, one traced from upstream state transitions and then tested with the same million [11]. The project's own framing is that fitting is a legitimate strategy supporting a different kind of confidence, not an illegitimate one [12].
Now the numbers, which are claims about the project's harness and nobody else's pipeline. Accepted evidence is 2,931,581,192 matching checks over 2,897,819 source records [3][2], so roughly 1,012 checks per structure [1]. That rules out a single canonical-string compare per molecule and fits the declared comparison list, which runs from exact molecular state and fingerprints through serialization behavior, operation composition, concurrent execution, energies, gradients and errors [5]. The distance-geometry work is counted separately, at more than 2.75 billion matrix entries traversed [4], so the two totals are not additive. And 3,968 shard tasks across 31 phases is exactly 128 tasks per phase [2]. A round number that lands on a power of two is a scheduler, not a coincidence.
For the 2.93 billion to mean anything about your workload, three things have to hold. Your calls have to sit inside the declared parity boundary, whose exact scope the article defers to VALIDATION.md rather than reproducing [5]. Your inputs have to resemble what the ChEMBL 37 structure table actually contains, because that is the corpus that was traversed [1]. And you have to accept the project's triage vocabulary: "consolidated accepted evidence" and "zero blocking mismatch" are counts after a classification step the project controls [3], and the article does not enumerate what a non-blocking mismatch looks like.
The pin is where I would put the budget question. Parity is asserted against RDKit 2026.03.1 specifically [1]. A fitted implementation responds to an upstream release by rerunning its corpus. A source-derived port has to re-read the upstream diff first, because the thing it claims to reproduce is the older revision's state transitions [6]. That is a real cost, paid per release, in exchange for knowing why the outputs agree instead of only that they did on 2.9 million inputs. In my context, if the parity boundary covers the calls my pipeline makes, that is the trade I would take.
